[j-nsp] QoS when there is no congestion

[Ross Halliday]

On Nov 17, 2016, at 8:07 AM, Jason Healy <jhealy at logn.net> wrote:
> 
> We're actually starting to experience this now.  We have a QFX with 10g and 1g links, and seeing significant drops on the 1g interfaces when traffic arrives on a 10g.  We have QoS enabled on our WAN connection, but haven't put it on the internal interfaces (where the drops are).
> 
> Do you have links to guidelines for basic QoS guidelines where you have this mismatched interface speed?  I haven't had a lot of luck finding out things like underlying buffer size, so that makes troubleshooting a little tricky.


Well, my experience has been with specific Cisco models, and I'm basing things I do based on that (lesson I learned: don't count on things to "just work"). If you do a search for QoS for iSCSI on the 3750 you will find a plethora of posts on the cisco-nsp list. What actually got me on to that was the documentation from our SAN vendor. Obviously we're kind of comparing apples to oranges here but it's all fruit, right? 

Funny, now that you mention it, I went and checked our EX4550s with 10G VMware hosts to our old 1G SAN... output drops! Nothing worse than 1259 out of 1868189159 for the iSCSI queue, but apparently I need to update my strategy. Since it's just the SAN on it I probably don't need so much buffer space allocated to the queue I created for vMotion etc (2011112229 packets overall).


On Nov 17, 2016, at 3:54 PM, Aaron <aaron1 at gvtc.com> wrote:
>
> Hmmmm, interestingly, I'm getting word from someone that is running test
traffic through my network that they are seeing out of sequence with small
test frames (64 or 68 bytes)....but any frames larger than that are fine.
Your mention of "reordering" gives me concern.  I wonder how/where I would
solve that.

Sorry Aaron, I probably should have used another word. I was thinking of jitter when I wrote that. Outright reordering/resequencing would be interesting to see - I don't have enough experience in that department to comment.


- Ross